Real-Time Industrial Defect Detection on Edge Hardware Using Fine-Tuned YOLOv8: A Systematic Benchmark on the NEU Surface Defect Database and MVTec AD with Automotive & Battery Manufacturing Extensions

Emmanuel Ezeji Somtochukwu; Nitesh Rijal

arxiv: 2606.07659 · v1 · pith:VOG2OVR5 · submitted 2026-06-03 · cs.CV · eess.IV

Real-Time Industrial Defect Detection on Edge Hardware Using Fine-Tuned YOLOv8: A Systematic Benchmark on the NEU Surface Defect Database and MVTec AD with Automotive & Battery Manufacturing Extensions

Emmanuel Ezeji Somtochukwu , Nitesh Rijal This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-28 06:27 UTCgrok-4.3pith:VOG2OVR5 record.json open to challenge →

classification cs.CV eess.IV

keywords defect detectionYOLOv8edge computingindustrial inspectionreal-time detectionNEU datasetMVTec ADTensorRT

0 comments

The pith

Industrial-YOLO, a fine-tuned YOLOv8, exceeds 120 FPS on Jetson Orin hardware at 98.5 percent mAP on industrial defect datasets.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Industrial-YOLO as an edge-optimized framework based on fine-tuned YOLOv8 for detecting surface defects in manufacturing. It benchmarks the system on the NEU steel defect database, the MVTec AD dataset, and custom extensions for automotive and battery manufacturing that include scratches, pits, and inclusions. Target-specific optimizations using TensorRT and OpenVINO are applied to meet edge hardware constraints. The central demonstration is that these steps produce inference speeds above 120 FPS on the NVIDIA Jetson Orin while holding mean average precision at 98.5 percent. This setup is shown to support direct deployment on active assembly lines for automated optical inspection.

Core claim

Industrial-YOLO, built on a fine-tuned YOLOv8 architecture with TensorRT and OpenVINO optimizations, achieves inference speeds exceeding 120 FPS on the NVIDIA Jetson Orin platform while maintaining a mean Average Precision of 98.5 percent across the NEU surface defect database, MVTec AD, and custom automotive manufacturing extensions.

What carries the argument

Fine-tuned YOLOv8 architecture accelerated by TensorRT and OpenVINO for edge deployment on NVIDIA Jetson Orin.

If this is right

The framework supports zero-latency defect detection directly on active automotive assembly lines.
It provides a scalable approach for next-generation automated optical inspection systems in manufacturing.
High-speed performance holds across steel sheet defects and structural anomalies such as scratches, pits, and inclusions.
The same optimization pipeline can extend the approach to battery manufacturing datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar edge optimizations could be applied to other YOLO variants or detection architectures on comparable hardware.
Testing on additional manufacturing lines beyond automotive would reveal how dataset-specific the accuracy holds.
Further hardware generations might push speeds higher while preserving the same accuracy level.

Load-bearing premise

The fine-tuning process and TensorRT/OpenVINO optimizations preserve the reported accuracy on the target industrial datasets without post-hoc data selection or hardware-specific retraining.

What would settle it

Running the reported model and optimizations on the Jetson Orin with the NEU and MVTec test sets yields either under 120 FPS or under 98.5 percent mAP.

read the original abstract

Automated surface defect detection is critical for ensuring rigorous quality control in high-speed manufacturing environments. While deep learning models offer remarkable accuracy, deploying them on resource-constrained edge hardware without introducing significant latency remains a persistent challenge. This paper presents Industrial-YOLO, an edge-optimized framework built upon a fine-tuned YOLOv8 architecture specifically engineered for real-time industrial defect detection. We conduct a systematic benchmark utilizing the NEU surface defect database for steel sheets and the MVTec AD dataset, supplemented with custom automotive manufacturing extensions representing real-world structural anomalies (scratches, pits, and inclusions). To bridge the gap between algorithmic complexity and edge hardware constraints, target-specific optimizations are introduced via TensorRT and OpenVINO acceleration engines. Experimental results demonstrate that Industrial-YOLO achieves a high-velocity inference speed exceeding 120 FPS on the NVIDIA Jetson Orin platform while maintaining an exceptional mean Average Precision (mAP) of 98.5%. The proposed framework showcases highly robust, zero-latency performance when deployed directly onto an active automotive assembly line, offering a scalable blueprint for next-generation automated optical inspection (AOI) systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This is a routine fine-tuning of YOLOv8 plus standard TensorRT/OpenVINO acceleration on known defect datasets, with headline numbers that lack the controls needed to verify them.

read the letter

The main takeaway is that the paper applies an existing model to industrial defect detection and reports strong speed and accuracy on edge hardware, but the supporting details are missing.

What it does is benchmark YOLOv8 after fine-tuning on the NEU steel dataset and MVTec AD, then add custom examples from automotive and battery lines. The optimizations target the Jetson Orin and combine FPS with mAP, which at least addresses the practical constraints of real-time inspection.

The soft spots are more central. The abstract gives no training splits, hyperparameter choices, or validation protocol. It also does not confirm that the 98.5% mAP was measured after the TensorRT and OpenVINO steps rather than before, or that the same resolution and test set were used for both accuracy and speed. Without those checks, the central claim is difficult to evaluate. There are also no error bars or ablation results on how the custom data or the acceleration steps affect performance.

This paper is aimed at engineers who want a worked example of deploying object detection for manufacturing quality control. Readers looking for new methods or rigorous comparisons will not find much.

I would not bring this to a reading group. I would not cite it. It does not deserve peer review because the evidence for the main performance claims is too thin to assess.

Referee Report

2 major / 0 minor

Summary. The paper introduces Industrial-YOLO, a fine-tuned YOLOv8 architecture with TensorRT and OpenVINO optimizations for real-time surface defect detection on edge hardware. It reports systematic benchmarks on the NEU surface defect database, MVTec AD dataset, and custom automotive/battery manufacturing extensions, claiming inference exceeding 120 FPS on NVIDIA Jetson Orin while achieving 98.5% mAP.

Significance. If the performance numbers are shown to hold under standard controls (identical splits for accuracy and speed, mAP measured on the final optimized model), the work would supply a practical reference implementation for AOI systems in manufacturing. The combination of public benchmarks with domain-specific extensions is a positive aspect.

major comments (2)

[Abstract] Abstract: the headline claim of 98.5% mAP together with >120 FPS on the optimized Jetson Orin model cannot be evaluated because the text supplies no information on training/validation splits, hyperparameter selection, or whether mAP was computed on the TensorRT/OpenVINO model versus the unoptimized fine-tune.
[Experimental Results] The central claim requires an explicit side-by-side comparison of mAP (and any other accuracy metrics) before and after TensorRT/OpenVINO optimization on the identical held-out test splits used for the reported 98.5% figure; without this comparison the preservation of accuracy under acceleration remains unverified.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for greater experimental transparency. We will revise the manuscript to address the points raised.

read point-by-point responses

Referee: [Abstract] Abstract: the headline claim of 98.5% mAP together with >120 FPS on the optimized Jetson Orin model cannot be evaluated because the text supplies no information on training/validation splits, hyperparameter selection, or whether mAP was computed on the TensorRT/OpenVINO model versus the unoptimized fine-tune.

Authors: We agree the manuscript should explicitly state these details. The revised version will update the abstract and add a methods subsection specifying the train/validation/test splits, hyperparameter selection process, and confirming that the 98.5% mAP was evaluated on the final TensorRT/OpenVINO-optimized model. revision: yes
Referee: [Experimental Results] The central claim requires an explicit side-by-side comparison of mAP (and any other accuracy metrics) before and after TensorRT/OpenVINO optimization on the identical held-out test splits used for the reported 98.5% figure; without this comparison the preservation of accuracy under acceleration remains unverified.

Authors: We will add the requested side-by-side comparison in the experimental results section, reporting mAP and related metrics on the identical held-out test splits before and after optimization to verify accuracy preservation. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical benchmark with no derivation chain

full rationale

The paper presents an empirical study of fine-tuned YOLOv8 with TensorRT/OpenVINO optimizations, reporting measured FPS and mAP values on NEU, MVTec AD, and custom datasets. No equations, first-principles derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. All headline figures are direct experimental outcomes on external benchmarks, so the work contains no self-referential reduction of results to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated in the provided text. The central claim rests on empirical benchmark numbers whose supporting details are absent.

pith-pipeline@v0.9.1-grok · 5753 in / 1217 out tokens · 37984 ms · 2026-06-28T06:27:09.214706+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 1 canonical work pages

[1]

Jocher, G., Chaurasia, A., & Qiu, J. (2023). Ultralytics YOLOv8. GitHub. https://github.com/ultralytics/ultralytics

2023
[2]

Krizhevsky, A., et al. (2012). ImageNet classification with deep CNNs. NeurIPS, 25, 1097–1105

2012
[3]

Simonyan, K., & Zisserman, A. (2015). Very deep CNNs for large-scale image recognition. ICLR

2015
[4]

He, K., et al. (2016). Deep residual learning for image recognition. CVPR, 770–778

2016
[5]

Redmon, J., et al. (2016). You only look once. CVPR, 779–788

2016
[6]

Jocher, G., et al. (2020). YOLOv5. GitHub

2020
[7]

Liu, W., et al. (2016). SSD: Single shot multibox detector. ECCV, 21–37

2016
[8]

Tao, X., et al. (2018). Automatic metallic surface defect detection with CNNs. Applied Sciences, 8(9),

2018
[9]

Niu, S., et al. (2020). Defect image sample generation with GAN. IEEE Trans. Autom. Sci. Eng., 17(3), 1611–1622

2020
[10]

Bergmann, P., et al. (2019). MVTec AD: A dataset for unsupervised anomaly detection. CVPR, 9592–9600

2019
[11]

Roth, K., et al. (2022). Towards total recall in industrial anomaly detection. CVPR, 14310–14320

2022
[12]

Yu, J., et al. (2021). FastFlow: Unsupervised anomaly detection via 2D normalizing flows. arXiv:2111.07677

work page arXiv 2021
[13]

Zhao, R., et al. (2019). Deep learning for machine health monitoring. Mech. Syst. Signal Process., 115, 213–237

2019
[14]

Tremblay, J., et al. (2018). Deep object pose estimation for semantic robotic grasping. CoRL

2018
[15]

Lee, J., et al. (2020). Industrial AI for Industry 4.0. Manufacturing Letters, 18, 20–23

2020
[16]

NVIDIA. (2023). TensorRT 8.6 Developer Guide

2023
[17]

Zhao, T., et al. (2021). Real-time object detection on embedded systems. J. Real-Time Image Process., 18(4)

2021
[18]

Shi, Y., et al. (2016). Automatic road crack detection. IEEE Trans. ITS, 17(12), 3434–3445

2016
[19]

Bacioiu, D., et al. (2019). Automated defect classification of TIG welding. NDT&E International,

2019
[20]

Bauer, J., et al. (2020). Industrial paint coating defect detection. SPIE, 11421

2020
[21]

Wieler, M., & Hahn, T. (2007). Weakly supervised learning for industrial inspection. DAGM

2007
[22]

Lin, T.Y., et al. (2014). Microsoft COCO. ECCV, 740–755. [NEW] Song, K., & Yan, Y. (2013). A noise robust method based on CLBP for hot-rolled steel strip surface defects. Applied Surface Science, 285, 858–864. Emmanuel Ezeji Somtochukwu is founder of Zema AI Labs (Abuja, Nigeria) and MEng candidate in Intelligent Manufacturing and Control Engineering at Z...

2014

[1] [1]

Jocher, G., Chaurasia, A., & Qiu, J. (2023). Ultralytics YOLOv8. GitHub. https://github.com/ultralytics/ultralytics

2023

[2] [2]

Krizhevsky, A., et al. (2012). ImageNet classification with deep CNNs. NeurIPS, 25, 1097–1105

2012

[3] [3]

Simonyan, K., & Zisserman, A. (2015). Very deep CNNs for large-scale image recognition. ICLR

2015

[4] [4]

He, K., et al. (2016). Deep residual learning for image recognition. CVPR, 770–778

2016

[5] [5]

Redmon, J., et al. (2016). You only look once. CVPR, 779–788

2016

[6] [6]

Jocher, G., et al. (2020). YOLOv5. GitHub

2020

[7] [7]

Liu, W., et al. (2016). SSD: Single shot multibox detector. ECCV, 21–37

2016

[8] [8]

Tao, X., et al. (2018). Automatic metallic surface defect detection with CNNs. Applied Sciences, 8(9),

2018

[9] [9]

Niu, S., et al. (2020). Defect image sample generation with GAN. IEEE Trans. Autom. Sci. Eng., 17(3), 1611–1622

2020

[10] [10]

Bergmann, P., et al. (2019). MVTec AD: A dataset for unsupervised anomaly detection. CVPR, 9592–9600

2019

[11] [11]

Roth, K., et al. (2022). Towards total recall in industrial anomaly detection. CVPR, 14310–14320

2022

[12] [12]

Yu, J., et al. (2021). FastFlow: Unsupervised anomaly detection via 2D normalizing flows. arXiv:2111.07677

work page arXiv 2021

[13] [13]

Zhao, R., et al. (2019). Deep learning for machine health monitoring. Mech. Syst. Signal Process., 115, 213–237

2019

[14] [14]

Tremblay, J., et al. (2018). Deep object pose estimation for semantic robotic grasping. CoRL

2018

[15] [15]

Lee, J., et al. (2020). Industrial AI for Industry 4.0. Manufacturing Letters, 18, 20–23

2020

[16] [16]

NVIDIA. (2023). TensorRT 8.6 Developer Guide

2023

[17] [17]

Zhao, T., et al. (2021). Real-time object detection on embedded systems. J. Real-Time Image Process., 18(4)

2021

[18] [18]

Shi, Y., et al. (2016). Automatic road crack detection. IEEE Trans. ITS, 17(12), 3434–3445

2016

[19] [19]

Bacioiu, D., et al. (2019). Automated defect classification of TIG welding. NDT&E International,

2019

[20] [20]

Bauer, J., et al. (2020). Industrial paint coating defect detection. SPIE, 11421

2020

[21] [21]

Wieler, M., & Hahn, T. (2007). Weakly supervised learning for industrial inspection. DAGM

2007

[22] [22]

Lin, T.Y., et al. (2014). Microsoft COCO. ECCV, 740–755. [NEW] Song, K., & Yan, Y. (2013). A noise robust method based on CLBP for hot-rolled steel strip surface defects. Applied Surface Science, 285, 858–864. Emmanuel Ezeji Somtochukwu is founder of Zema AI Labs (Abuja, Nigeria) and MEng candidate in Intelligent Manufacturing and Control Engineering at Z...

2014